Investigating fine-tuning versus zero-shot learning for general large language models when predicting cancer survival from initial oncology consultation documents

Thu, 01 Jan 2026 00:00:00 +0000

This study investigates whether fine-tuning open-weight large language models (LLMs) improves cancer survival prediction from oncology consultation notes compared to zero-shot approaches. Using Meta’s Llama models on 59,800 patient records from BC Cancer, fine-tuning consistently improved performance over zero-shot inference, but did not outperform smaller task-specific NLP models. The findings suggest both approaches merit continued investigation, with the best choice depending on clinical context and practical constraints such as hardware, privacy, and deployment feasibility.

T. Phaterpekar | Nunez Lab for Computational Mental Health & Cancer Care

Investigating fine-tuning versus zero-shot learning for general large language models when predicting cancer survival from initial oncology consultation documents